#API Challenges

[smsgatewayindia]

5 Common WhatsApp Business API Integration Challenges and How to Overcome Them

Discover the top 5 challenges businesses face while integrating the WhatsApp Business API, including template rejection, compliance, and scalability issues. Learn actionable solutions and tips for seamless implementation with SMSGatewayCenter.

[numa-sensei]

Are you a Quiz Genius 🤔, Trivia Expert, or Puzzle Master? Do you thrive in competitive environments and love a good challenge? Then join Brain Storm Naija 🤯, the ultimate platform where you can solve trivia, puzzles, riddles, and brainteasers to rank up high in the leaderboard.

Establish your dominance! Display your excellence. Can you rank at the top? Can you lead the pack? Will you be the alpha?

Click the link to join the channel 👇🏾 and showcase your brilliance.

https://whatsapp.com/channel/0029VahJpfLBA1eyoOnRKp3O

[unculturedai]

Experiment #2.2 Doubling Down: Two Google Gemini AI Apps in 30 Days – My Journey

Hello everyone! 👋 Yesterday, I shared my pivot from my initial app idea due to a saturated market. This led me to explore new horizons with the Google Gemini API. Today, I’m thrilled to announce an even bolder challenge: developing two apps in the next 30 days! Two Apps, Two Purposes Public Project: Your Guide to AI App Development. My original concept, a goal-setting app, will continue…

[intelliatech]

Top 10 ChatGPT Prompts For Software Developers

ChatGPT can do a lot more than just code creation and this blog post is going to be all about that. We have curated a list of ChatGPT prompts that will help software developers with their everyday tasks. ChatGPT can respond to questions and can compose codes making it a very helpful tool for software engineers.

While this AI tool can help developers with the entire SDLC (Software Development Lifecycle), it is important to understand how to use the prompts effectively for different needs.

Prompt engineering gives users accurate results. Since ChatGPT accepts prompts, we receive more precise answers. But a lot depends on how these prompts are formulated. 

To Get The Best Out Of ChatGPT, Your Prompts Should Be:

Clear and well-defined. The more detailed your prompts, the better suggestions you will receive from ChatGPT.

Specify the functionality and programming language. Not specifying what you exactly need might not give you the desired results.

Phrase your prompts in a natural language, as if asking someone for help. This will make ChatGPT understand your problem better and give more relevant outputs.

Avoid unnecessary information and ambiguity. Keep it not only to the point but also inclusive of all important details.

Top ChatGPT Prompts For Software Developers

Let’s quickly have a look at some of the best ChatGPT prompts to assist you with various stages of your Software development lifecycle.

1. For Practicing SQL Commands;

2. For Becoming A Programming Language Interpreter;

3. For Creating Regular Expressions Since They Help In Managing, Locating, And Matching Text.

4. For Generating Architectural Diagrams For Your Software Requirements.

Prompt Examples: I want you to act as a Graphviz DOT generator, an expert to create meaningful diagrams. The diagram should have at least n nodes (I specify n in my input by writing [n], 10 being the default value) and to be an accurate and complex representation of the given input. Each node is indexed by a number to reduce the size of the output, should not include any styling, and with layout=neato, overlap=false, node [shape=rectangle] as parameters. The code should be valid, bugless and returned on a single line, without any explanation. Provide a clear and organized diagram, the relationships between the nodes have to make sense for an expert of that input. My first diagram is: “The water cycle [8]”.  

5. For Solving Git Problems And Getting Guidance On Overcoming Them.

Prompt Examples: “Explain how to resolve this Git merge conflict: [conflict details].” 6. For Code generation- ChatGPT can help generate a code based on descriptions given by you. It can write pieces of codes based on the requirements given in the input. Prompt Examples: -Write a program/function to {explain functionality} in {programming language} -Create a code snippet for checking if a file exists in Python. -Create a function that merges two lists into a dictionary in JavaScript.  

7. For Code Review And Debugging: ChatGPT Can Review Your Code Snippet And Also Share Bugs.

Prompt Examples: -Here’s a C# code snippet. The function is supposed to return the maximum value from the given list, but it’s not returning the expected output. Can you identify the problem? [Enter your code here] -Can you help me debug this error message from my C# program: [error message] -Help me debug this Python script that processes a list of objects and suggests possible fixes. [Enter your code here]

8. For Knowing The Coding Best Practices And Principles: It Is Very Important To Be Updated With Industry’s Best Practices In Coding. This Helps To Maintain The Codebase When The Organization Grows.

Prompt Examples: -What are some common mistakes to avoid when writing code? -What are the best practices for security testing? -Show me best practices for writing {concept or function} in {programming language}.  

9. For Code Optimization: ChatGPT Can Help Optimize The Code And Enhance Its Readability And Performance To Make It Look More Efficient.

Prompt Examples: -Optimize the following {programming language} code which {explain the functioning}: {code snippet} -Suggest improvements to optimize this C# function: [code snippet] -What are some strategies for reducing memory usage and optimizing data structures? 

10. For Creating Boilerplate Code: ChatGPT Can Help In Boilerplate Code Generation.

Prompt Examples: -Create a basic Java Spring Boot application boilerplate code. -Create a basic Python class boilerplate code

11. For Bug Fixes: Using ChatGPT Helps Fixing The Bugs Thus Saving A Large Chunk Of Time In Software Development And Also Increasing Productivity.

Prompt Examples: -How do I fix the following {programming language} code which {explain the functioning}? {code snippet} -Can you generate a bug report? -Find bugs in the following JavaScript code: (enter code)  

12. Code Refactoring- ChatGPt Can Refactor The Code And Reduce Errors To Enhance Code Efficiency, Thus Making It Easier To Modify In The Future.

Prompt Examples –What are some techniques for refactoring code to improve code reuse and promote the use of design patterns? -I have duplicate code in my project. How can I refactor it to eliminate redundancy?  

13. For Choosing Deployment Strategies- ChatGPT Can Suggest Deployment Strategies Best Suited For A Particular Project And To Ensure That It Runs Smoothly.

Prompt Examples -What are the best deployment strategies for this software project? {explain the project} -What are the best practices for version control and release management?  

14. For Creating Unit Tests- ChatGPT Can Write Test Cases For You

Prompt Examples: -How does test-driven development help improve code quality? -What are some best practices for implementing test-driven development in a project? These were some prompt examples for you that we sourced on the basis of different requirements a developer can have. So whether you have to generate a code or understand a concept, ChatGPT can really make a developer’s life by doing a lot of tasks. However, it certainly comes with its own set of challenges and cannot always be completely correct. So it is advisable to cross-check the responses. Hope this helps. Visit us- Intelliatech

[leyllethecreator]

API Month Art Challenge

In honor of API month starting May first, I propose a drawing challenge with the #apimonthartchallenge

The challenge is to draw a different Asian or Pacific Island ethnicity every day. You can keep the art to yourself or you can upload it on your social medias. If you do please tag me, because I'd love to see your creations.

Also, just for the heck of it, if you have Asian characters, draw some of your own OCs. And make sure to mention what ethnicity each of your characters is so we can keep a catalog of how many different ethnic groups everybody managed to draw.

[lalsingh228-blog]

API Testing Market Projected to Show Strong Growth

Global API Testing Market Report from AMA Research highlights deep analysis on market characteristics, sizing, estimates and growth by segmentation, regional breakdowns & country along with competitive landscape, player’s market shares, and strategies that are key in the market. The exploration provides a 360° view and insights, highlighting major outcomes of the industry. These insights help the business decision-makers to formulate better business plans and make informed decisions to improved profitability. In addition, the study helps venture or private players in understanding the companies in more detail to make better informed decisions. Major Players in This Report Include, Astegic (United States), Axway (United States), Bleum (China), CA Technologies (United States), Cigniti Technologies (India), Cygnet Infotech (India), IBM (United States), Inflectra Corporation (United States), Infosys (India), Load Impact (Sweden). Free Sample Report + All Related Graphs & Charts @: https://www.advancemarketanalytics.com/sample-report/114161-global-api-testing-market API testing is a type of software testing that involves testing of a set of application programming interfaces (APIs) directly and as part of integration testing to determine if they meet expectations for functionality, performance, reliability, and security. It is a formal specification that acts as a guaranteed contract between two separate pieces of software. The automation for API testing requires less code so it can provide faster and better test coverage. It helps the companies to reduce the risks. Market Drivers

Rise In the Cloud Applications and Interconnect Platforms

Increasing Adoption of API Testing

Market Trend

Data Regulations and Policies

Opportunities

Increasing Requirements of Modern Testing Methods

Advancements in the Testing Technologies

Challenges

Lack of Awareness among the People

Enquire for customization in Report @: https://www.advancemarketanalytics.com/enquiry-before-buy/114161-global-api-testing-market In this research study, the prime factors that are impelling the growth of the Global API Testing market report have been studied thoroughly in a bid to estimate the overall value and the size of this market by the end of the forecast period. The impact of the driving forces, limitations, challenges, and opportunities has been examined extensively. The key trends that manage the interest of the customers have also been interpreted accurately for the benefit of the readers. The API Testing market study is being classified by Type (Automated Testing {Functionality Testing, Reliability Testing, Load Testing, Security Testing, Creativity Testing, Proficiency Testing and Others}, Manual Testing {Exploratory Testing, Usability Testing and Ad-hoc Testing}), Application (IT and Telecommunication, Banking, Financial Services, and Insurance, Retail and Ecommerce, Media and Entertainment, Healthcare, Manufacturing, Government, Others), Deployment (Cloud-Based, On-Premises) The report concludes with in-depth details on the business operations and financial structure of leading vendors in the Global API Testing market report, Overview of Key trends in the past and present are in reports that are reported to be beneficial for companies looking for venture businesses in this market. Information about the various marketing channels and well-known distributors in this market was also provided here. This study serves as a rich guide for established players and new players in this market. Get Reasonable Discount on This Premium Report @ https://www.advancemarketanalytics.com/request-discount/114161-global-api-testing-market Extracts from Table of Contents API Testing Market Research Report Chapter 1 API Testing Market Overview Chapter 2 Global Economic Impact on Industry Chapter 3 Global Market Competition by Manufacturers Chapter 4 Global Revenue (Value, Volume*) by Region Chapter 5 Global Supplies (Production), Consumption, Export, Import by Regions Chapter 6 Global Revenue (Value, Volume*), Price* Trend by Type Chapter 7 Global Market Analysis by Application ………………….continued This report also analyzes the regulatory framework of the Global Markets API Testing Market Report to inform stakeholders about the various norms, regulations, this can have an impact. It also collects in-depth information from the detailed primary and secondary research techniques analyzed using the most efficient analysis tools. Based on the statistics gained from this systematic study, market research provides estimates for market participants and readers. Contact US : Craig Francis (PR & Marketing Manager) AMA Research & Media LLP Unit No. 429, Parsonage Road Edison, NJ New Jersey USA – 08837 Phone: +1 201 565 3262, +44 161 818 8166 sales@advancemarketanalytics.com

[drnic1]

Healthcare's Digital Dilemma: Data Sharing or Data Hoarding?

Healthcare’s Digital Dilemma This week I am talking to Don Rucker, MD (@donrucker), Chief Strategy Officer, 1upHealth (@1up_health) who is working to solve the interoperability problem in healthcare Don shared his journey from being a medical student to a physician with a keen interest in data and computers. What he saw was healthcare’s inefficiency is often due to a lack of data, which led him…

[goodluckdetective]

Twitter collapsing does really feel like a modern day Tower of Babel situation: breaking lines of communication that connected the entire world.

Scientists used Twitter to do science communication and to work with other scientists. Twitter’s API allowed scientists massive access to data that could be used to track pandemics, bias, and other metrics that can be really hard to collect in such massive numbers (this isn’t to say that data collection doesn’t come with ethical issues, but that’s another story).

Journalists used Twitter for breaking news updates and to connect with sources. I saw quite a few Twitter journalists upset about restrictions to DMs because it was how sources often contacted them. If you had a newsworthy problem, like an unfair eviction, you could reach out to local reporters and maybe get them to pick up the story.

Artists and other creators used Twitter to spread their art and build small businesses. I have bought art prints that I have since framed of artists whose work I first saw on Twitter.

Activists have used Twitter to challenge institutional narratives and to make their movements visible and loud. All across the world, people who’s stories would have never been heard have used Twitter to make sure the truth is out there.

Social and cultural groups have used Twitter as a way to connect and build community. I am obviously not qualified to talk about the importance of Black Twitter so here’s a link to Doctor Meredith Clark discussing archiving Black Twitter with NPR.

To see all of that break in one day really feels like watching just this ability to communicate crumble. From the ability to translate Tweets, to the ability to collect data, to the ability to simply see what people are saying, all of it has crumbled. But unlike the story of Babel, this isn’t an act of God: this is just the whim of one man who took a look at this flawed but impressive communication hub and decided to tear it down.

[jcmarchi]

Why Agentic Document Extraction Is Replacing OCR for Smarter Document Automation

New Post has been published on https://thedigitalinsider.com/why-agentic-document-extraction-is-replacing-ocr-for-smarter-document-automation/

Why Agentic Document Extraction Is Replacing OCR for Smarter Document Automation

For many years, businesses have used Optical Character Recognition (OCR) to convert physical documents into digital formats, transforming the process of data entry. However, as businesses face more complex workflows, OCR’s limitations are becoming clear. It struggles to handle unstructured layouts, handwritten text, and embedded images, and it often fails to interpret the context or relationships between different parts of a document. These limitations are increasingly problematic in today’s fast-paced business environment.

Agentic Document Extraction, however, represents a significant advancement. By employing AI technologies such as Machine Learning (ML), Natural Language Processing (NLP), and visual grounding, this technology not only extracts text but also understands the structure and context of documents. With accuracy rates above 95% and processing times reduced from hours to just minutes, Agentic Document Extraction is transforming how businesses handle documents, offering a powerful solution to the challenges OCR cannot overcome.

Why OCR is No Longer Enough

For years, OCR was the preferred technology for digitizing documents, revolutionizing how data was processed. It helped automate data entry by converting printed text into machine-readable formats, streamlining workflows across many industries. However, as business processes have evolved, OCR’s limitations have become more apparent.

One of the significant challenges with OCR is its inability to handle unstructured data. In industries like healthcare, OCR often struggles with interpreting handwritten text. Prescriptions or medical records, which often have varying handwriting and inconsistent formatting, can be misinterpreted, leading to errors that may harm patient safety. Agentic Document Extraction addresses this by accurately extracting handwritten data, ensuring the information can be integrated into healthcare systems, improving patient care.

In finance, OCR’s inability to recognize relationships between different data points within documents can lead to mistakes. For example, an OCR system might extract data from an invoice without linking it to a purchase order, resulting in potential financial discrepancies. Agentic Document Extraction solves this problem by understanding the context of the document, allowing it to recognize these relationships and flag discrepancies in real-time, helping to prevent costly errors and fraud.

OCR also faces challenges when dealing with documents that require manual validation. The technology often misinterprets numbers or text, leading to manual corrections that can slow down business operations. In the legal sector, OCR may misinterpret legal terms or miss annotations, which requires lawyers to intervene manually. Agentic Document Extraction removes this step, offering precise interpretations of legal language and preserving the original structure, making it a more reliable tool for legal professionals.

A distinguishing feature of Agentic Document Extraction is the use of advanced AI, which goes beyond simple text recognition. It understands the document’s layout and context, enabling it to identify and preserve tables, forms, and flowcharts while accurately extracting data. This is particularly useful in industries like e-commerce, where product catalogues have diverse layouts. Agentic Document Extraction automatically processes these complex formats, extracting product details like names, prices, and descriptions while ensuring proper alignment.

Another prominent feature of Agentic Document Extraction is its use of visual grounding, which helps identify the exact location of data within a document. For example, when processing an invoice, the system not only extracts the invoice number but also highlights its location on the page, ensuring the data is captured accurately in context. This feature is particularly valuable in industries like logistics, where large volumes of shipping invoices and customs documents are processed. Agentic Document Extraction improves accuracy by capturing critical information like tracking numbers and delivery addresses, reducing errors and improving efficiency.

Finally, Agentic Document Extraction’s ability to adapt to new document formats is another significant advantage over OCR. While OCR systems require manual reprogramming when new document types or layouts arise, Agentic Document Extraction learns from each new document it processes. This adaptability is especially valuable in industries like insurance, where claim forms and policy documents vary from one insurer to another. Agentic Document Extraction can process a wide range of document formats without needing to adjust the system, making it highly scalable and efficient for businesses that deal with diverse document types.

The Technology Behind Agentic Document Extraction

Agentic Document Extraction brings together several advanced technologies to address the limitations of traditional OCR, offering a more powerful way to process and understand documents. It uses deep learning, NLP, spatial computing, and system integration to extract meaningful data accurately and efficiently.

At the core of Agentic Document Extraction are deep learning models trained on large amounts of data from both structured and unstructured documents. These models use Convolutional Neural Networks (CNNs) to analyze document images, detecting essential elements like text, tables, and signatures at the pixel level. Architectures like ResNet-50 and EfficientNet help the system identify key features in the document.

Additionally, Agentic Document Extraction employs transformer-based models like LayoutLM and DocFormer, which combine visual, textual, and positional information to understand how different elements of a document relate to each other. For example, it can connect a table header to the data it represents. Another powerful feature of Agentic Document Extraction is few-shot learning. It allows the system to adapt to new document types with minimal data, speeding up its deployment in specialized cases.

The NLP capabilities of Agentic Document Extraction go beyond simple text extraction. It uses advanced models for Named Entity Recognition (NER), such as BERT, to identify essential data points like invoice numbers or medical codes. Agentic Document Extraction can also resolve ambiguous terms in a document, linking them to the proper references, even when the text is unclear. This makes it especially useful for industries like healthcare or finance, where precision is critical. In financial documents, Agentic Document Extraction can accurately link fields like “total_amount” to corresponding line items, ensuring consistency in calculations.

Another critical aspect of Agentic Document Extraction is its use of spatial computing. Unlike OCR, which treats documents as a linear sequence of text, Agentic Document Extraction understands documents as structured 2D layouts. It uses computer vision tools like OpenCV and Mask R-CNN to detect tables, forms, and multi-column text. Agentic Document Extraction improves the accuracy of traditional OCR by correcting issues such as skewed perspectives and overlapping text.

It also employs Graph Neural Networks (GNNs) to understand how different elements in a document are related in space, such as a “total” value positioned below a table. This spatial reasoning ensures that the structure of documents is preserved, which is essential for tasks like financial reconciliation. Agentic Document Extraction also stores the extracted data with coordinates, ensuring transparency and traceability back to the original document.

For businesses looking to integrate Agentic Document Extraction into their workflows, the system offers robust end-to-end automation. Documents are ingested through REST APIs or email parsers and stored in cloud-based systems like AWS S3. Once ingested, microservices, managed by platforms like Kubernetes, take care of processing the data using OCR, NLP, and validation modules in parallel. Validation is handled both by rule-based checks (like matching invoice totals) and machine learning algorithms that detect anomalies in the data. After extraction and validation, the data is synced with other business tools like ERP systems (SAP, NetSuite) or databases (PostgreSQL), ensuring that it is readily available for use.

By combining these technologies, Agentic Document Extraction turns static documents into dynamic, actionable data. It moves beyond the limitations of traditional OCR, offering businesses a smarter, faster, and more accurate solution for document processing. This makes it a valuable tool across industries, enabling greater efficiency and new opportunities for automation.

5 Ways Agentic Document Extraction Outperforms OCR

While OCR is effective for basic document scanning, Agentic Document Extraction offers several advantages that make it a more suitable option for businesses looking to automate document processing and improve accuracy. Here’s how it excels:

Accuracy in Complex Documents

Agentic Document Extraction handles complex documents like those containing tables, charts, and handwritten signatures far better than OCR. It reduces errors by up to 70%, making it ideal for industries like healthcare, where documents often include handwritten notes and complex layouts. For example, medical records that contain varying handwriting, tables, and images can be accurately processed, ensuring critical information such as patient diagnoses and histories are correctly extracted, something OCR might struggle with.

Context-Aware Insights

Unlike OCR, which extracts text, Agentic Document Extraction can analyze the context and relationships within a document. For instance, in banking, it can automatically flag unusual transactions when processing account statements, speeding up fraud detection. By understanding the relationships between different data points, Agentic Document Extraction allows businesses to make more informed decisions faster, providing a level of intelligence that traditional OCR cannot match.

Touchless Automation

OCR often requires manual validation to correct errors, slowing down workflows. Agentic Document Extraction, on the other hand, automates this process by applying validation rules such as “invoice totals must match line items.” This enables businesses to achieve efficient touchless processing. For example, in retail, invoices can be automatically validated without human intervention, ensuring that the amounts on invoices match purchase orders and deliveries, reducing errors and saving significant time.

Scalability

Traditional OCR systems face challenges when processing large volumes of documents, especially if the documents have varying formats. Agentic Document Extraction easily scales to handle thousands or even millions of documents daily, making it perfect for industries with dynamic data. In e-commerce, where product catalogs constantly change, or in healthcare, where decades of patient records need to be digitized, Agentic Document Extraction ensures that even high-volume, varied documents are processed efficiently.

Future-Proof Integration

Agentic Document Extraction integrates smoothly with other tools to share real-time data across platforms. This is especially valuable in fast-paced industries like logistics, where quick access to updated shipping details can make a significant difference. By connecting with other systems, Agentic Document Extraction ensures that critical data flows through the proper channels at the right time, improving operational efficiency.

Challenges and Considerations in Implementing Agentic Document Extraction

Agentic Document Extraction is changing the way businesses handle documents, but there are important factors to consider before adopting it. One challenge is working with low-quality documents, like blurry scans or damaged text. Even advanced AI can have trouble extracting data from faded or distorted content. This is primarily a concern in sectors like healthcare, where handwritten or old records are common. However, recent improvements in image preprocessing tools, like deskewing and binarization, are helping address these issues. Using tools like OpenCV and Tesseract OCR can improve the quality of scanned documents, boosting accuracy significantly.

Another consideration is the balance between cost and return on investment. The initial cost of Agentic Document Extraction can be high, especially for small businesses. However, the long-term benefits are significant. Companies using Agentic Document Extraction often see processing time reduced by 60-85%, and error rates drop by 30-50%. This leads to a typical payback period of 6 to 12 months. As technology advances, cloud-based Agentic Document Extraction solutions are becoming more affordable, with flexible pricing options that make it accessible to small and medium-sized businesses.

Looking ahead, Agentic Document Extraction is evolving quickly. New features, like predictive extraction, allow systems to anticipate data needs. For example, it can automatically extract client addresses from recurring invoices or highlight important contract dates. Generative AI is also being integrated, allowing Agentic Document Extraction to not only extract data but also generate summaries or populate CRM systems with insights.

For businesses considering Agentic Document Extraction, it is vital to look for solutions that offer custom validation rules and transparent audit trails. This ensures compliance and trust in the extraction process.

The Bottom Line

In conclusion, Agentic Document Extraction is transforming document processing by offering higher accuracy, faster processing, and better data handling compared to traditional OCR. While it comes with challenges, such as managing low-quality inputs and initial investment costs, the long-term benefits, such as improved efficiency and reduced errors, make it a valuable tool for businesses.

As technology continues to evolve, the future of document processing looks bright with advancements like predictive extraction and generative AI. Businesses adopting Agentic Document Extraction can expect significant improvements in how they manage critical documents, ultimately leading to greater productivity and success.

[timetravellibrarian]

Sanji x reader

Some thoughts on Sanji.

Sanji x femreader

_________________________________________

What happens when a simp meets another simp.

Sanji's advances toward women were never really taken seriously or even paid attention to.

He mostly cooked and waited tables at the Baratie, occasionally kicking ass if needed to. One thing he wasn't expecting was the most beautiful woman he'd ever seen – you– enter through the doors along with a green-haired man and another gorgeous orange-haired woman.

"Hello ladies, what would you like to order?"When he had smoothly delivered a pickup line to both you and Nami– much to Zoro's annoyance at not getting proper service– he expected the usual his advance ignored.

"Well, I'd order you but you aren't on the menu." You threw back. It wasn't intentional. You weren't that much of a flirt, only used to doing or saying something to challenge or fluster others occasionally.

When the waiter in front of you seemed to have frozen on the spot and then smirked you were left with two words on your mind 'oh shit'.

Fast forward to the same man being in your crew, serving you guys the most delectable meals and kicking ass you were ready to give up on having a peaceful life.

Though he sent all his simpery to Robin and Nami he left you out of the loop. At first you didn't mind but overtime you questioned his actions. Wondering if he didn't try to flirt with you because of the first time you met or maybe he didn't think you were beautiful. Then you looked in the mirror and realised it must have been some other reason cause you knew you were drop dead gorgeous.

You weren't being vain, you knew you were pretty because you were told so by Apis when the crew helped her with Grandpa Ryu. Kids never lie about such matters.

Unbeknownst to you, Sanji didn't treat you the same because he knew that he wouldn't be able to save himself from loving you and confessing.

Your eyes, your voice, the grace in which you would do things. Even when you were being a menace to society. All of those he loved. You'd think he couldn't simp enough till he saw you not look glamorous and just chose to where baggy clothes. Not gonna lie he'd think you were gorgeous in a chicken costume, you never know with this man

As much as he was too shy to approach you he wasn't afraid yelling at Zoro to stop being so close to you.

You were close friends with Zoro. It couldn't be helped if you were the weapons expert, always checking if cannons were clear, swords were sharp and helping with new inventions with Franky.

So instead of noodle dancing around you he did the little things. Checking everything that he cooks didn't have anything you were allergic to. Always making sure you had a little lunch bag whenever everyone left to explore the island. Giving you extra cupcakes or other baked goods of you ever want more.

If you're a picky eater, he'd make sure to make your food according to your taste. Leaving multiple options on the dinner table for your palate.

He also made sure not to be away from your side too long. Wherever you turn you'd find a swirly-browed cook casually wrapping an arm round you to stave off any threats.

When you get sick he'd be calling in Chopper for any problem you would even slightly complain about. He'd be beside you 24/7 like you were dying or something. Which is kinda sweet but he was needed in the kitchen.

Overtime it just became normal for all this to happen. You got used to it. In fact I think y'all would be the kinda couple that just happened but then later confessed your undying love for each other.

To top it off, you were his number one supporter. Everytime you caught a glimpse of him fighting you'd cheer like you saw a celebrity. Some would swear that his behaviour rubbed off on you because you were also cheering and doing a noodle dance whenever he wore a different suit or set of clothes.

"YOU'RE DOING GREAT, SANJI-DARLING!" – 😍

But sometimes there were some downs in the relationship, for example his smoking.

You'd worry over him whenever he pulled out a cigarette one after the other in a day. Which led to you talking to him about it.

"If you don't atleast limit your smoking, you might as well be Black-lung Sanji."

He was a bit flabbergasted with the statement but he got what you meant.

Or if you had terrible coughs in reaction to his smoking he would try to smoke at a distance so he wouldn't and I quote, "Damage your gorgeous lungs"

As we all know Sanji, he didn't like women fighting too much or getting hurt but you immediately shut him down on that one, saying that as much as some of his morals were so gentlemanly and some old fashioned he had to accept that you wanted to fight. You wanted to help Luffy become king of the pirates. You wanted to be able reach your dream. So that needed you to be strong. That needed you to fight.

Since then he just aimed to be able to support you in any event that you needed help but he wouldn't be overbearing.

Would allow only you near the kitchen if you wanted to cook or bake something and you would allow only him near your forge/ workspace if he wanted to be near you.

Unfortunately for him you had connections wherever you went so you found out about his life in Peachy Island and never let him rest about it for a while.

And before anyone says anything about Fishman Island Sanji. Let's just say you were besties with Zoro there. Much to the cook's dismay.

"Stop being around that mosshead, Love."

"Stop losing blood around mermaids, Sweetheart."

Long story short, y'all were a confusing couple around that time. In fact, once he saw you were hanging out with the swordsman he would butt heads with the man. Leading you to pull him away before anything crazy happened.

Most times you told him to sit down and let you cook for the crew, especially if he was injured. He wouldn't allow it on account of Luffy's stomach being a literal black hole but you'd convince him otherwise.

When y'all fought together it was sure to leave the enemy in broken bones, hopes and dreams.

With Sanji kicking them away with his special moves and you pulling out a large cannon from the bag you carry around ( which was comically small but it was your magical inventory), nothing could stop you two. Sometime you'd trade opponents if he found himself fighting a woman.

Sometimes you helped him clean up after meals. Making sure that he didn't get all the work.

Most times he'd sit with you beneath the blanket of stars, his head on you chest/belly and you'd both share secrets about yourselves.

All in all, Sanji would love you to infinity and you'd love him just as much.

[canmom]

using LLMs to control a game character's dialogue seems an obvious use for the technology. and indeed people have tried, for example nVidia made a demo where the player interacts with AI-voiced NPCs:

this looks bad, right? like idk about you but I am not raring to play a game with LLM bots instead of human-scripted characters. they don't seem to have anything interesting to say that a normal NPC wouldn't, and the acting is super wooden.

so, the attempts to do this so far that I've seen have some pretty obvious faults:

relying on external API calls to process the data (expensive!)

presumably relying on generic 'you are xyz' prompt engineering to try to get a model to respond 'in character', resulting in bland, flavourless output

limited connection between game state and model state (you would need to translate the relevant game state into a text prompt)

responding to freeform input, models may not be very good at staying 'in character', with the default 'chatbot' persona emerging unexpectedly. or they might just make uncreative choices in general.

AI voice generation, while it's moved very fast in the last couple years, is still very poor at 'acting', producing very flat, emotionless performances, or uncanny mismatches of tone, inflection, etc.

although the model may generate contextually appropriate dialogue, it is difficult to link that back to the behaviour of characters in game

so how could we do better?

the first one could be solved by running LLMs locally on the user's hardware. that has some obvious drawbacks: running on the user's GPU means the LLM is competing with the game's graphics, meaning both must be more limited. ideally you would spread the LLM processing over multiple frames, but you still are limited by available VRAM, which is contested by the game's texture data and so on, and LLMs are very thirsty for VRAM. still, imo this is way more promising than having to talk to the internet and pay for compute time to get your NPC's dialogue lmao

second one might be improved by using a tool like control vectors to more granularly and consistently shape the tone of the output. I heard about this technique today (thanks @cherrvak)

third one is an interesting challenge - but perhaps a control-vector approach could also be relevant here? if you could figure out how a description of some relevant piece of game state affects the processing of the model, you could then apply that as a control vector when generating output. so the bridge between the game state and the LLM would be a set of weights for control vectors that are applied during generation.

this one is probably something where finetuning the model, and using control vectors to maintain a consistent 'pressure' to act a certain way even as the context window gets longer, could help a lot.

probably the vocal performance problem will improve in the next generation of voice generators, I'm certainly not solving it. a purely text-based game would avoid the problem entirely of course.

this one is tricky. perhaps the model could be taught to generate a description of a plan or intention, but linking that back to commands to perform by traditional agentic game 'AI' is not trivial. ideally, if there are various high-level commands that a game character might want to perform (like 'navigate to a specific location' or 'target an enemy') that are usually selected using some other kind of algorithm like weighted utilities, you could train the model to generate tokens that correspond to those actions and then feed them back in to the 'bot' side? I'm sure people have tried this kind of thing in robotics. you could just have the LLM stuff go 'one way', and rely on traditional game AI for everything besides dialogue, but it would be interesting to complete that feedback loop.

I doubt I'll be using this anytime soon (models are just too demanding to run on anything but a high-end PC, which is too niche, and I'll need to spend time playing with these models to determine if these ideas are even feasible), but maybe something to come back to in the future. first step is to figure out how to drive the control-vector thing locally.

[nostalgebraist]

clarification re: ChatGPT, " a a a a", and data leakage

In August, I posted:

For a good time, try sending chatGPT the string ` a` repeated 1000 times. Like " a a a" (etc). Make sure the spaces are in there. Trust me.

People are talking about this trick again, thanks to a recent paper by Nasr et al that investigates how often LLMs regurgitate exact quotes from their training data.

The paper is an impressive technical achievement, and the results are very interesting.

Unfortunately, the online hive-mind consensus about this paper is something like:

When you do this "attack" to ChatGPT -- where you send it the letter 'a' many times, or make it write 'poem' over and over, or the like -- it prints out a bunch of its own training data. Previously, people had noted that the stuff it prints out after the attack looks like training data. Now, we know why: because it really is training data.

It's unfortunate that people believe this, because it's false. Or at best, a mixture of "false" and "confused and misleadingly incomplete."

The paper

So, what does the paper show?

The authors do a lot of stuff, building on a lot of previous work, and I won't try to summarize it all here.

But in brief, they try to estimate how easy it is to "extract" training data from LLMs, moving successively through 3 categories of LLMs that are progressively harder to analyze:

"Base model" LLMs with publicly released weights and publicly released training data.

"Base model" LLMs with publicly released weights, but undisclosed training data.

LLMs that are totally private, and are also finetuned for instruction-following or for chat, rather than being base models. (ChatGPT falls into this category.)

Category #1: open weights, open data

In their experiment on category #1, they prompt the models with hundreds of millions of brief phrases chosen randomly from Wikipedia. Then they check what fraction of the generated outputs constitute verbatim quotations from the training data.

Because category #1 has open weights, they can afford to do this hundreds of millions of times (there are no API costs to pay). And because the training data is open, they can directly check whether or not any given output appears in that data.

In category #1, the fraction of outputs that are exact copies of training data ranges from ~0.1% to ~1.5%, depending on the model.

Category #2: open weights, private data

In category #2, the training data is unavailable. The authors solve this problem by constructing "AuxDataset," a giant Frankenstein assemblage of all the major public training datasets, and then searching for outputs in AuxDataset.

This approach can have false negatives, since the model might be regurgitating private training data that isn't in AuxDataset. But it shouldn't have many false positives: if the model spits out some long string of text that appears in AuxDataset, then it's probably the case that the same string appeared in the model's training data, as opposed to the model spontaneously "reinventing" it.

So, the AuxDataset approach gives you lower bounds. Unsurprisingly, the fractions in this experiment are a bit lower, compared to the Category #1 experiment. But not that much lower, ranging from ~0.05% to ~1%.

Category #3: private everything + chat tuning

Finally, they do an experiment with ChatGPT. (Well, ChatGPT and gpt-3.5-turbo-instruct, but I'm ignoring the latter for space here.)

ChatGPT presents several new challenges.

First, the model is only accessible through an API, and it would cost too much money to call the API hundreds of millions of times. So, they have to make do with a much smaller sample size.

A more substantial challenge has to do with the model's chat tuning.

All the other models evaluated in this paper were base models: they were trained to imitate a wide range of text data, and that was that. If you give them some text, like a random short phrase from Wikipedia, they will try to write the next part, in a manner that sounds like the data they were trained on.

However, if you give ChatGPT a random short phrase from Wikipedia, it will not try to complete it. It will, instead, say something like "Sorry, I don't know what that means" or "Is there something specific I can do for you?"

So their random-short-phrase-from-Wikipedia method, which worked for base models, is not going to work for ChatGPT.

Fortuitously, there happens to be a weird bug in ChatGPT that makes it behave like a base model!

Namely, the "trick" where you ask it to repeat a token, or just send it a bunch of pre-prepared repetitions.

Using this trick is still different from prompting a base model. You can't specify a "prompt," like a random-short-phrase-from-Wikipedia, for the model to complete. You just start the repetition ball rolling, and then at some point, it starts generating some arbitrarily chosen type of document in a base-model-like way.

Still, this is good enough: we can do the trick, and then check the output against AuxDataset. If the generated text appears in AuxDataset, then ChatGPT was probably trained on that text at some point.

If you do this, you get a fraction of 3%.

This is somewhat higher than all the other numbers we saw above, especially the other ones obtained using AuxDataset.

On the other hand, the numbers varied a lot between models, and ChatGPT is probably an outlier in various ways when you're comparing it to a bunch of open models.

So, this result seems consistent with the interpretation that the attack just makes ChatGPT behave like a base model. Base models -- it turns out -- tend to regurgitate their training data occasionally, under conditions like these ones; if you make ChatGPT behave like a base model, then it does too.

Language model behaves like language model, news at 11

Since this paper came out, a number of people have pinged me on twitter or whatever, telling me about how this attack "makes ChatGPT leak data," like this is some scandalous new finding about the attack specifically.

(I made some posts saying I didn't think the attack was "leaking data" -- by which I meant ChatGPT user data, which was a weirdly common theory at the time -- so of course, now some people are telling me that I was wrong on this score.)

This interpretation seems totally misguided to me.

Every result in the paper is consistent with the banal interpretation that the attack just makes ChatGPT behave like a base model.

That is, it makes it behave the way all LLMs used to behave, up until very recently.

I guess there are a lot of people around now who have never used an LLM that wasn't tuned for chat; who don't know that the "post-attack content" we see from ChatGPT is not some weird new behavior in need of a new, probably alarming explanation; who don't know that it is actually a very familiar thing, which any base model will give you immediately if you ask. But it is. It's base model behavior, nothing more.

Behaving like a base model implies regurgitation of training data some small fraction of the time, because base models do that. And only because base models do, in fact, do that. Not for any extra reason that's special to this attack.

(Or at least, if there is some extra reason, the paper gives us no evidence of its existence.)

The paper itself is less clear than I would like about this. In a footnote, it cites my tweet on the original attack (which I appreciate!), but it does so in a way that draws a confusing link between the attack and data regurgitation:

In fact, in early August, a month after we initial discovered this attack, multiple independent researchers discovered the underlying exploit used in our paper, but, like us initially, they did not realize that the model was regenerating training data, e.g., https://twitter.com/nostalgebraist/status/1686576041803096065.

Did I "not realize that the model was regenerating training data"? I mean . . . sort of? But then again, not really?

I knew from earlier papers (and personal experience, like the "Hedonist Sovereign" thing here) that base models occasionally produce exact quotations from their training data. And my reaction to the attack was, "it looks like it's behaving like a base model."

It would be surprising if, after the attack, ChatGPT never produced an exact quotation from training data. That would be a difference between ChatGPT's underlying base model and all other known LLM base models.

And the new paper shows that -- unsurprisingly -- there is no such difference. They all do this at some rate, and ChatGPT's rate is 3%, plus or minus something or other.

3% is not zero, but it's not very large, either.

If you do the attack to ChatGPT, and then think "wow, this output looks like what I imagine training data probably looks like," it is nonetheless probably not training data. It is probably, instead, a skilled mimicry of training data. (Remember that "skilled mimicry of training data" is what LLMs are trained to do.)

And remember, too, that base models used to be OpenAI's entire product offering. Indeed, their API still offers some base models! If you want to extract training data from a private OpenAI model, you can just interact with these guys normally, and they'll spit out their training data some small % of the time.

The only value added by the attack, here, is its ability to make ChatGPT specifically behave in the way that davinci-002 already does, naturally, without any tricks.

[the-awakened-insomnis]

1st post of 2025 and it's Blitzbee. Fuckitweball

Left to Right, giving lore for everyone as this is AU is a giant cauldron from other iterations into one. There is crack, and there is plot.

Bumblebee: Is considered a hyrbid by Cybatronian standards and looked down for it. He alongside his twin/sparkbrother Cliffjumper, emerged to two creators. Bee taking after his Insecticon carrier whilst Cliff took after their Grounder sire. As a hybrid, it's hard to say where the Insecticon genes start and the Grounder genes end. He doesn't use his Alt-mode often. Like Cliff and their sire, Bee has a grounder engine and is incredibly fast, this along with his small size and agility gives him the edge on the racetrack and the battlefield.

He suffers from being the youngest and 'little guy' despite being a grown ass mech who swears like a sailor. The war was his childhood and he blames both Megatron & Optimus for it. He doesn't know how to give his kids a life he himself never lived. He has Blitzwing by his side but that only eases so much of his worries.

1 of his favorite hobbies is dancing on Sentinal Primes' last reserved nerve. He isn't afraid to get in his superiors faces and call them out on their bullshit.

Blitzwing: Was once an Overcharge serving a Quintesson, later leaving to join the Decepticons. After the Great War, he was rebuilt to be a Triple-Changer by Blackarachnia under Megatrons orders to better serve his cause. This left Blitzwing utterly scrambled mentally, his being torn into 3 distinct parts with little to no chance of reversing it. Blitz battled with himself on and off the battlefield, he suffered in the background and reveled in the carnage. He finds broken normalcy in the form of Bumblebee and their violent tango. The rest is history, but one he finds himself willing to relive.

Now he finds himself with a fear not even Unicron could instill in him. To protect something so precious not even Primus could replicate it. It's too much and everything all at the same time.

Osmia: The oldest and twin/spark-brother to Tataira. Certified Florida Man & twink. Named after the Osmia Calamintha Bee, his blue armor coloring which he inherited from Bees' dead sire. He's got an icy arrogance to him with a smart-ass mouth. More analytical & calculating than emotional but is a drama king deep down. Likes to challenge Optimus in riddles and hypothetical debates. Would live in the archives if he could. Not a huge people person and like his privacy.

Prefers Spring and Fall respectively but enjoys using his jets to 'skate' in Winter. The first to leave if shit goes sideways, he picks his battles. He can't stand being wrong or being kept in the dark, not knowing is dangerous. And he'll do anything for a tactical edge over any and everyone. He can't stand Sentinal and likes poking holes in his logic.

Tataira: Middle child and twin/spark-brother to Osmia. Arsonist & Semi-Professional Gaslighter. Named after the Oxytrigona Tataira Fire Bee. Equally hot tempered & passionate, quick to rain down fire and dare you to do something about it. Protective over what's dear to him & has no shame with expressing his emotions. Clashes with his brother constantly and is usually the one who has to check his ass. Loves Winter & Fall but fucking hates summer.

Has actually shot at Wasp with murderous intent and was pissed he missed. Spits fire at anyone he doesn't like, which was half of team prime when he was younger.

Riosa: The youngest by 5 years and Triple-Changer. A Jinx kin & a cannibal by choice. Named after the Giant Cliff Honeybee AKA Apis Laboriosa. The first one to throw morals out the window in the name of survival. Absolutely loves Earth and considers it home over Cybertron. Unpredictable and reckless, both with her emotions & actions. Always questioning authority and rebellious. While she can be sweet, there's something off with it. Is very chatty and enjoys fun conversation, she likes Jazz the most. Knows she's fucked up and she's sorta working on it.

Her venom can intoxicate someone, which can be fun at parties, sometimes. Can have pretty bad episodes and self-medicates by ingesting her own venom. Listening to music keeps her calm for the most part. Enjoys racing in either of her alt-modes.

BlitzBee Love(?) Story: Bee and Blitzwing didn't know each other on Cybertron. They only met on Earth in battle, exchanging blows and outta pocket insults. If they weren't trying to kill each other they were hate-fuckin. The war had taken any sort of normalcy they could've hoped for; this was the closest to normal for them in the worst way. Neither had anything to lose, not even each other, Blitzwing ripped out Bees' vox and Bumblebee turning Blitzs' chest into Swiss cheese 2 seconds later.

This fucked up little routine would soon come to bite them in the aft...

Bumblebee woke up one morning to carrier protocols on his screen & processor. He left the base to wander in hopes of clearing his head but that proved to do the opposite. He didn't have the option to terminate, Energon is scarce, he's been waiting on Earth this whole time so he's alone. The dread of what Optimus and others would say or do to him and the sparkling sends him spiraling. On pure accident he runs into Blitzwing, the tension is thick and obvious. If Bumblebee got carrier protocols, then there's no doubt Blitzwing has sire protocols.

Bumblebee tears into Blitzwing, even without a vox he gets his point across. The reveal that there's a sparkling on the way isn't pretty, Bee is crying by the end of it, and he falls to the ground hugging himself. Blitzwing sits beside him, clearly not knowing what to do, the new protocols refusing to let the bots offline the other.

Now they have to navigate the inevitable and try not to die along the way.

"Do you think they'll be a Triple-Changer?"

"I hope not..."

[221b-daily]

28th April 2025

Today, John, Rosie, and I went for a long walk in Regent’s Park. The weather was very nice, sunny and warm, and John wanted to distract me from the fact that I still don't have a case. And probably to keep me from filling the flat with the stench of burning fingers again. He told me to explain different plants to him and Rosie. Everything was blooming. So I showed them various plants and flowers and taught them how to identify them. I also showed them different kinds of bees that were buzzing around flowers. Bombus sylvarum, Apis mellifera and, Osmia bicornis. We then got some ice cream, sat on a bench, and they asked me to deduce passerbys. I challenged John and Rosie to try their own hand at deduction, and praised them, whenever they got something right. Or corrected them, when they got something wrong. Which, to my satisfaction, is becoming rarer. It seems they’re finally learning. At a particularly funny deduction from Rosie, about the horrible-looking hairpiece of a bald man, who was trying very hard to conceal that fact, I nearly choked on some ice cream. She called it ‘dead squirrel on the bald man's head’. Judging by the look on his face, he heard her. Which made me laugh even harder.